Predicting radiocephalic arteriovenous fistula success with machine learning

This research paper proposes a machine learning-based point-of-care tool to enhance the prediction of successful unassisted radiocephalic arteriovenous fistula (AVF) use. Using data from 704 patients, the study developed and validated various machine learning models (logistic, penalized logistic, decision tree, random forest, and boosted tree) using baseline clinical characteristics and ultrasound parameters. The Lasso model, demonstrating high performance and parsimony, was selected as the final model. This model incorporates outflow vein diameter, flow volume, and the absence of >50% luminal stenosis to predict successful AVF use within one year. An online calculator is available for clinical application.

Patrick Heindel, Tanujit Dey, Jessica D. Feliz, Dirk M. Hentschel, Deepak L. Bhatt, Mohammed Al-Omran, Michael Belkin, C. Keith Ozaki, Mohamad A. Hussain